Skip to main content

... LUCAS BOSCAINI

> Exploring artificial cognition and implications on software engineering and safety.

Stores

TFTVs skip fine-tuning entirely by mapping contrastive activation steering directions directly into rank-one weight updates using only forward-pass statistics.

Post-Training Free Task Vectors

TFTVs skip fine-tuning entirely by mapping contrastive activation steering directions directly into rank-one weight updates using only forward-pass statistics.

• 443 words • 3 min
NeurIPS LLM Safety

Talks